Visualizing Categorical Data with SAS and R: Exercises

Michael Friendly Ernest Kwan Cathy LaBrish Bin Sun

March 7, 2012

Getting started
The goal of these workshops is to give you some hands-on experience using the methods described in the lectures. For each lecture, the plan is to have you try to reproduce, and possibly extend, the analysis and visualization of one or two examples used in that lecture, with SAS or R (or both), as you are most comfortable with. In addition, there may also be one or two new examples to try. Some materials for these workshops are available online at http://datavis.ca/courses/VCD/. For reference, Table 1 lists the principal data sets that are used in the lectures. The SAS column gives the name of the .sas le containing the data and the name of the SAS data set created from this. The R column gives the name of the data set in R. An asterisk (*) marks the rst use of each data set. Table 1: Data sets used in the lectures Data set SAS R Federalist * Cholesterol-HD * Arthritis * Berkeley * Hair-Eye-Color * Distant-Vision * Sex-Fun * MS diagnosis * Suicide-Rates * Berkeley Hair-Eye-Color * Mental-Health-SES * Distant-Vision * Berkeley Arthritis Volunteering * Arrests * Arthritis Womens-Labor-Force * madison fat arthrit berkeley haireye vision sexfun msdiag suicide berkeley haireye mental vision berkeley arthrit cowles arrests arthrit wlfpart Federalist Arthritis UCBAdmissions HairEyeColor VisualAcuity SexualFun MSPatients Suicide UCBAdmissions HairEyeColor Mental [vcdExtra] VisualAcuity UCBAdmissions Arthritis Cowles [car] Arrests [effects] Arthritis Womenlf [car]

Part 1

SAS and R setups

Here are a few things you should know to get started using SAS and R. 1

SAS All of the SAS programs, data sets and macros are stored in N:/Vcd. Documentation for all of my macro programs is available at http://datavis.ca/sasmac/, mostly under the Categorical data menu item. To use a data set in one of the exercises, start with %include catdata(dataset) for example, %include catdata(madison); %include catdata(arthrit); *-- Federalist papers, may by Madison; *-- Arthritis treatment data;

SAS

The SAS macro programs are installed in a macros directory that is automatically searched when you invoke it with %macroname(...). For example: %include catdata(madison); *-- Federalist papers, may by Madison; %goodfit(data=madison, var=count, freq=blocks, dist=poisson); Within SAS, you can access the web help page for any macro with %webhelp(macroname);. For example, %webhelp(goodfit); For help on SAS procedures, you can use the Help menu in SAS, or access the complete SAS Online documentation (v. 9.1.3) via %webhelp(online); R All of the R programs and data sets should be pre-installed in R on the server, but some of them may be outdated. The rst time you start R, you should update the available packages and load the vcdExtra package to make available some extra workshop materials. update.packages(ask=FALSE) library(vcdExtra) Thereafter, when you start R, you should begin with library(vcdExtra) to load the vcdExtra package. This includes vcd plus some extra data sets. To load a data set from an R package, use the data() function. To get a description of any R data set or function, use help(), or the short-hand ?. data(Federalist) ?Federalist ?mosaic # load the data # read the help() # help for mosaic() R

Home-work
In the limited time available in the lab sessions, it is unlikely that you will be able to complete all the exercises or to master all of these techniques. To prot more fully, and apply them to your own data, you should try to practice more at home or at work. To this end, most of the materials for this course and the workshops are available at http://datavis.ca/courses/VCD/. From there you can download the SAS data sets and macros (in the archive vcdprog.zip) and the necessary R packages. vcdprog.zip 2

1
1.1

Introduction & overview

Discrete distributions

In this exercise, we will examine the use of diagnostic and graphic methods for tting discrete distributions to the one-way frequency tables shown in Table 2 and Table 3. Table 2: Number of occurrences (k ) and number of blocks of text (nk ) of the word may in essays written by Madison. SAS: madison; R: Federalist Occurrences (k ) Blocks (nk ) 0 156 1 63 2 29 3 8 4 4 5 1 6 1

Table 3: Number of males in N=6115 Saxony families with 12 children. SAS: saxony; R: Saxony k 0 1 2 3 4 5 6 7 8 9 10 11 12 nk 3 24 104 286 670 1033 1343 1112 829 478 181 45 7 For the Madison data, it might be supposed that number of occurrences of the word may follow a Poisson distribution. 1. Fit a Poisson distribution to the data. Do you conclude that this distribution ts well or badly? 2. Examine the departure between the observed and tted frequencies with a hanging rootogram. Is there a pattern to the departure? 3. Use an Ord plot to diagnose the form of the distribution.1 4. If necessary, t a dierent distribution and examine visually with a hanging rootogram. With SAS, you can begin these steps with statements like the following: %include catdata(madison); %goodfit(data=madison, var=count, freq=blocks, dist=poisson, out=fit); %rootgram(data=fit, var=count, obs=blocks); %ordplot(data=madison, count=count, freq=blocks); With R the following statements will get you started: data(Federalist) gf <- goodfit(Federalist, type=poisson) summary(gf) plot(gf) Ord_plot(Federalist, main=may in Federalist papers) For the Saxony data, it might be supposed that number of male children follow a Binomial distribution, perhaps also with p(M ale) = 0.5. Repeat the steps above using the Saxony data, with a Binomial distribution.
To judge whether a coecient is positive or negative a tolerance is used. Be careful with the conclusions from the Ord plot, as it implements just some simple heuristics!
1

Tasks & questions

SAS

1.2

Testing association

The Berkeley data set (Bickel et al., 1975) involves graduate school applicants to the six largest departments at University of California, Berkeley in 1973. The applicants are classied by Admit (admitted vs. rejected), Gender (male vs. female), and Dept (A to F). This data set forms a 3-way table (i.e., a 2 2 6 table), but in this exercise, we will rst look at the 2 2 table of Admit by Gender, collapsed over Dept. In other words, we will ignore ( temporarily) the department to which each applicant applied. In later lectures and other exercises we will examine the full 3-way table. This analysis is of interest because an association between Admit and Gender dierent rates of admissions for male and female applicants could be considered as evidence of gender bias. 1. Construct the two-way frequency table by collapsing over Dept, and test whether Admit and Gender are independent. 2. Find the proportion admitted for males and females and relate these to the marginal frequencies and expected frequencies under independence. An independent relationship between Admit and Gender in the sample implies that cell frequencies would be proportionate to the product of marginal frequencies. For example, 59.5% (2691/4526) of applicants are male, 61.2% (2771/4526) of applicants are rejected. 3. Do these frequencies suggest approximately equal admission rates in males and females? If not, try to describe the possible gender dierence in admission. What is a good statistic to summarize the relation between gender and admission?

Tasks & questions

In SAS, with a data set in frequency form, you can collapse over any variables simply by SAS omitting them from the tables statement in PROC FREQ. See berkeley-freq.sas berkeleyfreq.sas

%include catdata(berkeley); proc freq data=berkeley; weight freq; tables gender*admit / chisq plots=all; format admit admit.; run; For R, the data set UCBAdmissions is represented as a 3-way table in table form. Use margin.table() to collapse any such table over omitted dimensions. CrossTable() in the gmodels package has many options for displaying quantities of interest in contingency tables. data(UCBAdmissions) UCB.GA <-margin.table(UCBAdmissions, c(1,2)) chisq.test(UCB.GA) library(gmodels) CrossTable(UCB.GA, prop.t=FALSE, prop.r=FALSE, expected=TRUE) R

1.3

Stratied analysis

The marginal analysis of the Berkeley data in Section 1.2 is misleading, because (as it turns out) it falsely assumes that relation between Admit and Gender is the same for each department. Only if this were true, would the collapsed table represent the relation in all departments. 4

Here, you will use a stratied analysis to examine the relation between Admit and Gender controlling for department (equivalently: a test of conditional independence Admit and Gender given Gender), as opposed to ignoring department in the marginal analysis. In Lectures II and III we will see some visualizations of these data that help to explain the contradictory results. 1. Carry out stratied tests of the association between Admit and Gender controlling for Dept. 2. What do you conclude about whether there is a dierence in rate of admission in each of the departments? 3. Test for conditional independence of Admit and Gender given Dept. 4. For one 2 2 table, the odds ratio is a reasonable summary measure of association. For a collection of k such tables, the Breslow-Day 2 tests whether these are equal across strata. What do you conclude? Using SAS, PROC FREQ will carry out a stratied analysis when there are more than two factors given in the tables statement You get a separate page of output for each stratum, followed by overall summary tests of the relation between the two primary table variables, controlling for the strata. These are the tests and statistics of most interest here. %include catdata(berkeley); proc freq data=berkeley; weight freq; tables dept * gender * admit /chisq cmh; format dept dept. admit admit.; run; There is no single, simple equivalent of this non-parametric analysis in R. Tests of similar hypotheses can be carried out using loglinear models (loglm()) and generalized linear models (glm()), described in later lectures. Here is one simple way to test for association of Admit and Gender separately for each department. oddsratio() calculates and tests the (log) odds ratio for each stratum. for (dept in 1:6) {print(chisq.test(UCBAdmissions[,,dept]))} summary(oddsratio(UCBAdmissions)) As well, some tests related to this exercise can be done as follows. You should try to understand what hypothesis is tested by each one. mantelhaen.test(UCBAdmissions) woolf_test(UCBAdmissions) R

Tasks & questions

SAS

1.4

Ordinal factors

The Sex-Fun data set (Hout et al., 1987) involves the responses of 91 couples to the questionnaire item: Sex is fun for me and my partner : (a) never or occasionally, (b) fairly often, (c) very often, (d) almost always. The data can be regarded as a 4 4 frequency table, where both factors can be considered to be ordered. For today, we will just consider testing whether there is an association between the husbands and wifes rating, and how to take ordinality into account, as well as the relatively small sample size. 5

In this problem, we could arguably be more interested in assessing to what extent the ratings actually agree there could actually be a strong negative association between the Husbands and Wifes rating of sexual fun! Methods for assessing agreement will be discussed in the lecture for Part II. 1. Carry out a simple Pearson 2 test for association of Husband and Wife. Are there any reasons to question the validity of this test? 2. Carry out the analogous Cochran-Mantel-Haenszel tests. SAS reports three dierent tests. Which one is most appropriate here, where both variables are ordinal? 3. For relatively small total sample size, where the asymptotic distribution of the Pearson 2 statistic is in doubt, Fishers exact test provides an alternative. Do it. 4. Examine the association statistics between the ratings. What would be a reasonable value to take as a non-parametric measure of correlation between Husband and Wife ratings?2 For SAS, you can carry out all these steps as follows shown below. See the le sexfun-cmh.sas. %include catdata(sexfun); proc freq data=sexfun order=data; weight count; tables husband * wife / cmh chisq exact nocol norow; run; For comparison with the association statistics, note that you can also calculate the Pearson correlation between the ratings using PROC CORR: proc corr data=sexfun ; weight count; var husband wife; run; In R, there is no equivalent of the CMH option in PROC FREQ. But you can do the other steps as follows. See sexfun-chisq.R. sexfun-chisq.R library(vcd) data(SexualFun) SexualFun chisq.test(SexualFun) fisher.test(SexualFun) assocstats(SexualFun) R # # # # show the table general association exact test association statistics
Tasks & questions

SAS
sexfuncmh.sas

2 Hint: the Cochran-Mantel-Haenszel 2 CM H = (n 1)r when both variables are assigned integer scores. 2 The Phi coecient, = P /n, but for larger than 2 2 tables it has a range 0 min( r 1, c 1). Cramers V is a corrected scaling of so that 0 V 1, similar to r2 .

Two-way and n-way tables

In these exercises you will examine several of the methods for visualizing association in twoway and n-way tables. The goal is not simply to see if an association exists, but to understand the nature or pattern of association why or how the variables are associated.

2.1

2 2 tables

In 2 2 tables, there are several measures of association between the row and column variables, but one of the simplest is the odds ratio, , 0 , estimated in the sample as = n11 /n12 = n11 n22 n21 /n22 n12 n21 When there is no association, = 1 (or, equivalently, log() = 0). For a stratied 2 2 k table, another question of interest is whether the odds ratios are the same for all strata, 1 = 2 , . . . , k . In this exercise, you will use fourfold displays to examine these questions. 2.1.1 Cholesterol and heart disease
Tasks & questions

1. Test whether there is an association between cholesterol and heart disease using a fourfold display. 2. What evidence is shown in the fourfold display regarding whether the odds ratio diers signicantly from 1? 3. Describe how cholesterol and heart disease are related. In SAS, the data are in fat.sas in the catdata directory. This le also runs a proc freq to give answers to the rst question. %include catdata(fat); %ffold(data=fat, var=diet disease); In R, you can just create the fat data using matrix(). chisq.test(fat) will give an approximate answer to the rst question. fat <- matrix( c(6, 4, 2, 11), 2, 2) dimnames(fat) <- list(diet=c("LoChol", "HiChol"), disease=c("No", "Yes")) fourfold(t(fat)) R

SAS

2.1.2

Berkeley data
Tasks & questions

Here, you will re-examine the Berkeley admissions data considered in Section 1.2 and Section 1.3 using fourfold plots. 1. Assess visually whether there is an association between gender and admission, in the marginal table that collapses over department. 2. Assess visually the association between gender and admission separately in each department. What do you conclude? 3. Examine the question of whether the odds ratio of admission for males and females is the same for all departments. [Hint: See the printed output from the ffold macro, or use the woolf_test() in R.] 7

In SAS, the table macro can be used to collapse a frequency table, or to change numeric SAS variables into character labels, suitable for the ffold macro. The required SAS statements for this exercise are contained in berk-4fold.sas. The rst step looks like this: berk-4fold.sas %include catdata(berkeley); %table(data=berkeley, out=berk2, var=Admit Gender, weight=freq, order=data); %ffold(data=berk2, var=Admit Gender); In R, margin.table() collapses a table over dimensions not listed in the second argument. The R statements below are contained in berk-4fold.R. See the vignette("vcd-tutorial") berk-4fold.R tutorial, Section 3.4, for other tests and plots of these data. R UCB <- aperm(UCBAdmissions, c(2,1,3)) fourfold(margin.table(UCB, c(1, 2))) # two-way plot fourfold(UCB, mfrow=c(2,3)) woolf_test(UCB) # three-way plot

2.2

Sieve diagrams

The Distant-vision data records the classication of visual acuity in the left and right eyes for a large sample of employees of the Royal Ordnance factor in the UK during world War II. There are two samples one of men and one of women. Here we will just look at the female sample. We expect to nd a strong degree of association (and agreement) between the grade recorded for the two eyes. The more interesting question is the nature of the association. A sieve diagram is useful for this purpose. 1. Construct a sieve diagram for these data. 2. Make sure you understand the principal by which this diagram is constructed, e.g., why are the boxes for the cells of the table aligned in rows and columns. 3. Describe the patterns you see in the density of shading.

Tasks & questions

In SAS, the data are found in the le vision.sas, and contain the data set women, for the SAS female sample. See the le vision-sieve.sas for this and more. vision%include catdata(vision); proc print data=women; run; title Vision data: Women; %sieveplot(data=women, var=right left, filltype=obsp);
sieve.sas

For R, you will nd the VisualAcuity data in the vcd package. In R, subset() allows easy R selection from a data.frame, which you can use to select the female sample. See the le vision-sieve.R for this and more. vision-sieve.R library(vcd) women <- subset(VisualAcuity, gender=="female", select=-gender) structable(~right + left, data=women) sieve(Freq ~ right + left, data = women, gp=shading_Friendly, labeling=labeling_values)

2.3

Observer agreement

Westlund & Kurland (1953) reported data on the diagnosis of multiple sclerosis (MS): two samples of patients, one from Winnipeg and one from New Orleans, were each rated by two neurologists (one from each city) in four diagnostic categories: Certain, Probable, Possible, and Doubtful. The questions of interest here concern how well the neurologists agree on diagnosis for these two samples of patients. 1. Obtain and test Cohens for each of the samples, using both equal spacing (CicchettiAllison) weights and inverse-square (Fleiss-Cohen) weights. What do you conclude about the strength of agreement in the two samples? 2. Obtain agreement charts for each of the samples. Visually, how does the strength of agreement shown in these charts compare to the measures? 3. Is there any evidence of dierence between the two neurologists in their frequency of use of the diagnostic categories (marginal heterogeneity)? 4. Test whether the strength of agreement diers between the two samples (only in SAS).

Tasks & questions

For SAS, the data are contained in the le msdiag.sas, as a single data set (msdiag) with vari- SAS ables Patients, N_rating and W_rating. With PROC FREQ, a stratied analysis by patients gives you most of the information to answer the questions concerning agreement. For agreement plots, with the agreeplot macro you need to subset the data to select each sample. The statements below are contained in the le msdiag-agree.sas. msdiagagree.sas

%include catdata(msdiag); *-- Agreement, separately, and controlling for Patients; proc freq data=msdiag; tables patients * N_rating * W_rating / norow nocol nopct agree plots=all; test kappa; weight count; run; data Winnipeg; set msdiag; where (patients="Winnipeg"); %agreeplot(data=Winnipeg, var=N_rating W_rating, title=Winnipeg patients); data NewOrleans; set msdiag; where (patients="New Orleans"); %agreeplot(data=NewOrleans, var=N_rating W_rating, title=New Orleans patients); In R, the data are contained in MSPatients, a 4 4 2 table. Patient is the third dimension, so you can use subscripting on the table to select the two samples separately. The statements below are contained in the le msdiag-agree.R. msdiagagree.R

library(vcd) Kappa(MSPatients[,,1]) Kappa(MSPatients[,,2])

# use ,weights="Fleiss" for F-C weights

agreementplot(t(MSPatients[,,1]), main = "Winnipeg Patients") agreementplot(t(MSPatients[,,2]), main = "New Orleans Patients")

2.4

Correspondence analysis

Correspondence analysis provides one way to visualize association between variables, both for nominal and ordinal categories. Here, we examine the relation between mental health and partents SES. For a 46 table, the total Pearson 2 can be accounted for in min(r 1, c1) = 3 dimensions, but we hope for a simpler explanation. 1. Carry out a correspondence analysis of these data. From the printed output, how many dimensions seem necessary here? 2. Interpret the results of the graph in terms of the question of whether both mental and ses can be considered to be linearly spaced. For SAS, the statements that carry out these steps are contained in the le mental-ca.sas. %include catdata(mental); data mental; set mental; format mental mental. ses ses.; run; Use the corresp macro to carry out the analysis and produce the graph. *-- Using the corresp macro; %corresp(data=mental, tables = mental / ses, weight = count, htext=1.3); In SAS 9.3, an equivalent plot can be obtained directly with PROC CORRESP (via SAS ODS Graphics). *-- Using SAS 9.3 ODS Graphics; proc corresp short data=mental; tables mental, ses; weight count; run;

Tasks & questions

mental-ca.sas

SAS

For R, the statements that carry out these steps are contained in the le mental-ca.R. Begin mental-ca.R by creating a two-way table, then use ca() for numerical results and plot() to graph the R result. library(vcdExtra) library(ca) Mental.tab <- xtabs(Freq ~ mental+ses, data=Mental) ca(Mental.tab) plot(ca(Mental.tab)) title(xlab=Dim 1, ylab=Dim 2)

10

3
3.1

Mosaic displays and loglinear models

Mosaic displays web applet

A web application to t loglinear models and produce mosaic displays is available at http://datavis.ca/online/mosaics/. It is not as exible as PROC GENMOD in SAS or the combination of loglm() and mosaic() in R, but is easy to use and contains several sample data sets, including the Berkelely data. Moreover, you can use it from home with your own data, uploaded from a le, or entered into a form. 1. Navigate to the Mosaic Displays web page, select the Berkeley Admission Data from the Sample datasets list, and press SelectData . 2. To illustrate, we will t the model [AD][GD] of conditional independence (Exercise 3 in Section 3.2), as follows: (a) (b) (c) (d) (e) (f) In the Variable order box enter: gender admit dept. Select CONDIT for FitType (You may wish to explore the Analysis Options and Display Options help page.) Press GetData Interpret each display in relation to the 2 tests printed above. What happens if you click on a tile in a mosaic?

Tasks & questions

3.2

Three-way+ tables

In this exercise you will exmine tting several dierent loglinear models for the data on admissions to graduate programes in Berkeley. We will use the GLM approach here.3 The table variables are dept, gender and admit, abbreviated as D, G and A. Note that admission is considered the response variable here, while gender and department are explanatory, so any reasonable model must include the [GD] association between gender and department. 1. Fit the homogeneous association model [AD][GD][AG] as generalized linear model for log frequency, and obtain tests for each term as well as the test for residual association. Does it appear that there is an association between admission and gender? 2. Interpret the model what does each term mean? What does the test for residual association mean? 3. Now t the reduced model [AD][GD] of conditional independence of admission and gender given department, A G | D. Obtain the residuals from this model and display these in a mosaic plot. What do you conclude? 4. It appears that conditional independence ts well, except in Department A. Fit a model that allows an association of admission and gender only in Department A.

Tasks & questions

The following steps will be helpful with SAS. If you need more help, see the statements in the le berkeley-glm.sas. berkeleyglm.sas

For display purposes it is useful to apply format statements to use as value labels for the table variables.
3 As explained in the lecture, such models can be t directly as loglinear models for the data in contingency table form, or as generalized linear models (GLM) with a log link and a Poisson distribution for the frequencies. In SAS, this amounts to the dierence between PROC CATMOD vs. PROC GENMOD. In R, the corresponding functions are loglm() and glm().

SAS

11

%include catdata(berkeley); *-- Apply formats to use value labels for numeric variables; data berkeley; set berkeley; format dept dept. admit admit. gender $sex.; The model of homogeneous association model [AD][GD][AG] (no 3-way association) can be specied using "|" notation as shown below. *-- Fit model of homogeneous associations: [AD][GD][AG]; proc genmod data=berkeley; class dept gender admit; model freq = dept|gender|admit@2 / dist=poisson type3 wald; run; The step below shows the pattern to (a) t a model, (b) obtain residuals in an obstats data set and (c) produce a mosaic plot for the model. *-- Fit model of conditional independence: [AD][GD]; proc genmod data=berkeley; class dept gender admit; model freq = dept|gender dept|admit / dist=poisson obstats; ods output obstats=obstats; run; %mosaic(data=obstats, vorder=admit gender dept, count=freq, resid=streschi, cellfill=dev, split=H V, title=Model: [AdmitDept] [GenderDept]); You can t a model allowing an association only in department A by dening a dummy variable that causes these cells to t perfectly. Add this term to the model statementn PROC GENMOD and see what happens. data berkeley; set berkeley; dept1AG = (gender=F) * admit * (dept=1); In R, all of these models can be t using glm(); the loglm() function cannot handle models with special terms (like dept1AG), but makes it quite easy to produce mosaic displays. The R statements below are in the le berkeley-glm.R. In R model formulas, ()^2 means all main berkeleyglm.R eects and interactions up to order 2. library(vcd) data("UCBAdmissions") structable(Dept ~ Admit+Gender,UCBAdmissions) ## conditional independence in UCB admissions data berk.mod1 <- loglm(~ Dept * (Gender + Admit), data=UCBAdmissions) berk.mod1 mosaic(berk.mod1, gp=shading_Friendly) ## all two-way model berk.mod2 <-loglm(~(Admit+Dept+Gender)^2, data=UCBAdmissions) berk.mod2 mosaic(berk.mod2, gp=shading_Friendly)

12

# compare models anova(berk.mod1, berk.mod2)

3.3

Survival on the Titanic

These exercises examine the tting of various loglinear models to data about survival on the Titanic, a 4-way table giving the cross-classication of 2201 passengers and crew, according to Gender (G): M vs. F Age (A): Adult vs. Child Class (C): 1st, 2nd, 3rd, Crew Survival (S): Died vs. Survived

1. One slight complication here is that there are 8 cells with zero frequencies. Four of these (male and female children in 1st and 2nd class who died) should be considered sampling zeros, but 4 (children among the crew) should probably be considered structural zeros cells where data could not occur. In these analyses, you can treat these all as sampling zeros by adding a small number to each cell. 2. It is natural to consider Survival as the natural response variable, and the remaining variables as explanatory. Therefored, all models should include the high-order term among Age Gender and Class. Therefore, the minimal null model is [AGC][S], which asserts that survival is jointly independent of Age, Gender and Class. Fit this model, and obtain a mosaic plot. Interpret the pattern of the residuals in this mosaic plot. 3. Fit a main eects model for survival, [AGC][AS][GS][CS], that includes an association of survival with each of age gender and class. Is this an adequate t? What does the pattern of residuals tell you about remaining associations? 4. What model would you use to allow an interaction of Age and Gender in their eect on Survival? Fit this model as above, and obtain the mosaic plot. For SAS, the statements for this exercise are in the le titanic-loglin.sas, but you should try titanicto gure them out for yourself rst. The table variables are named sex, age, class, survive. loglin.sas SAS Adjusting for 0 cells. %include catdata(titanic); data titanic; set titanic; count = count + .5; Fitting and graphing the null model. Note how the high-order term [AGC] is specied using | notation. title [AGC][S]: Baseline model; proc genmod data=titanic; class age sex class survive; model count=age|sex|class survive / dist=poisson type3 obstats; ods output obstats=obstats; %mosaic(data=obstats, vorder=class sex age survive, resid=reschi, title=[AGC][S]: Baseline model); 13

In R, it is easiest here to use the loglm() function to t a model, and the plot() method for R loglinear objects to obtain the mosaic plot. The statements below are contained in the le titanic-loglin.R. titanicloglin.R

library(vcd) data(Titanic) Titanic <- Titanic + 0.5 # adjust for 0 cells titanic.mod1 <- loglm(~ (Class * Age * Sex) + Survived, data=Titanic) titanic.mod1 plot(titanic.mod1, main="Model [AGC][S]") titanic.mod2 <- loglm(~ (Class * Age * Sex) + Survived*(Class + Age + Sex), data=Titanic) titanic.mod2 plot(titanic.mod2, main="Model [AGC][AS][GS][CS]") titanic.mod3 <- loglm(~ (Class * Age * Sex) + Survived*(Class + Age * Sex), data=Titanic) titanic.mod3 plot(titanic.mod3, main="Model [AGC][AS][GS][CS][AGS]") # compare models anova(titanic.mod1, titanic.mod2, titanic.mod3, test="chisq")

3.4

Ordinal factors and structured associations

When some of the variables in a contingency table represent ordered categories, we can usually gain both power and parsimony by tting models that take ordinality into account. We gain power by testing a more specic hypothesis than just a general association; we gain parsimony by tting fewer parameters. 3.4.1 Mental health and SES

In this exercise, you will examine several models for Mental-Health-SES data described in the lecture. In this two-way table, parents SES and child mental health status are both ordered factors. Simple ways of handling ordinal variable involve assigning scores to the table categories, and the simplest cases are to use integer scores, either for the row variable (column eects model), the column variable (row eects model), or both (uniform association model). 1. Fit the independence model to these data. How well does it t? Make a mosaic plot showing departures (residuals) from independence. What do you see here that suggests a simpler description of the association based on ordinality? 2. Assign integer scores to both parents SES and child mental health status. Fit the model of uniform association. How well does it t, compared to the independence model? Make a mosaic plot showing residuals from this model, and compare with that for the independence model. 3. If you have time, try also tting the column eects model, using scores for just mental health status, and the row eects model, using scores for just parents SES.

Tasks & questions

14

For SAS, most of the program statements for this exercise are contained in the le ment- SAS gen2.sas. Here are a few explanatory notes: mentgen2.sas In PROC GENMOD, a given variable can be either categorical (a factor) when declared in a CLASS statement or numeric (a quantitative covariate), but not both. To allow a factor to be treated as numeric in an association term, simply make a copy of that variable under a dierent name: %include catdata(mental); *-- Create numeric variables for row & column effects; data mental; set mental; m_lin = mental; s_lin = ses; format mental mental. ses ses.; Fit the independence model as follows, obtaining an output obstats data set of tted values, residuals, etc. that can be passed to the mosaic macro. proc genmod data=mental; class mental ses; model count = mental ses / dist=poisson obstats residuals; title Independence; ods output obstats=obstats; Visualize the remaining associations (residuals) with the mosaic macro. %mosaic(data=obstats, vorder=Mental SES, resid=stresdev, title=Mental Impairment and SES: Independence, split=H V, cellfill=dev 0.5, htext=2.0); Include one or both of the numeric versions of the table variables in an interaction term. For example, the model of uniform association, with a linear linear association is t and visualized as follows: proc genmod data=mental; class mental ses; model count = mental ses m_lin*s_lin / dist=poisson obstats residuals; title Lin x Lin; ods output obstats=obstats; run; %mosaic(data=obstats, vorder=Mental SES, resid=stresdev, title=Linear x Linear, split=H V, cellfill=dev 0.5, htext=2.0); For R the steps are similar. We use the data set Mental in the vcdExtra package. These R models are t using glm(), and visualized using mosaic.glm() applied to the glm object. See the le mental-glm.R for these and other steps. mental-glm.R Independence model: library(vcdExtra) data(Mental) indep <- glm(Freq ~ mental+ses, family = poisson, data = Mental) mosaic(indep,residuals_type="rstandard", labeling=labeling_residuals) 15

 Other models: Row eects, col eects and uniform association Create numeric equivalents of the table variables, then t models with interactions using the numeric scores. Cscore <- as.numeric(Mental$ses) Rscore <- as.numeric(Mental$mental) # row effects model (mental) roweff <- coleff <- glm(Freq ~ mental + ses + mental:Cscore, family = poisson, data = Mental) linlin <- glm(Freq ~ mental + ses + Rscore:Cscore, family = poisson, data = Mental) Visualize a given model with mosaic() as shown above for the independence model. 3.4.2 Distant vision: quasi-independence, quasi-symmetry

In Section 2.2 we examined the Distant-Vision data for women with a sieve diagram. Here, we will consider several specialized models that range between the independence model and the saturated model. The model of quasi-independence ignores the diagonal cells. The symmetry model tests whether the pattern of association is symmetric around the diagonal cells. Quasi-symmetry tests for symmetry, while allowing for dierences in the marginal frequencies. 1. Fit the models of independence, and quasi-independence ignoring the diagonal cells. Compare the goodness of t of the models numerically and with mosaic displays. 2. Fit the models of symmetry and quasi-symmetry (symmetric association, but allowing for dierences in marginal frequencies). Compare the goodness of t of the models numerically and with mosaic displays.

Tasks & questions

In SAS, recall that the data set women has numeric factors, left and right for the eye grades of the two eyes. The key to tting these models is to assign new variables corresponding to SAS the diagonal cells diag, and to the diagonally symmetric ones diag. The statements below are included in the le vision-quasi.sas, along with others not shown here. visionquasi.sas

%include catdata(vision); data women; set women; if left=right then do; diag=left; symmetry=-left; end; else do; diag=0; symmetry = abs(left-right); end; Independence and quasi-independence: The pattern is (a) t the model with PROC GENMOD, (b) obtain the obstats data set with residuals, (c) plot the data and residuals from the model with the mosaic macro. 16

title Independence model (women); proc genmod data=women; class Right Left; model Count = Left Right / dist=poisson link=log obstats residuals type3; ods exclude obstats; ods output obstats=obstats; run; %mosaic(data=obstats, vorder=Right Left, split=H V, htext=1.9, resid=reschi, title=%str(Independence, G2(9)=6671.5)); title Quasi-independence model (women); proc genmod data=women; class Right Left diag; model Count = Left Right diag / dist=poisson link=log obstats residuals type3; ods exclude obstats; ods output obstats=obstats; run; %mosaic(data=obstats, vorder=Right Left, split=H V, htext=1.9, resid=reschi, title=%str(Quasi-Independence, G2(5)=199.1)); Symmetry and quasi-symmetry: Use model Count = symmetry for strict symmetry and model Left Right symmetry quasi symmetry in the above. In R, you can t these models with glm() or gnm() in the gnm package. The gnm package provides special functions, Diag(), Symm() and others that considerably extend the range of R loglinear models for structured tables. See the vignette, Generalized nonlinear models in R, accessible in R as vignette("gnmOverview") for more details. At present mosaic() doesnt quite work with all of these more general models. The statements below are included in the le vision-quasi.R. vision-quasi.R library(vcdExtra) library(gnm) women <- subset(VisualAcuity, gender=="female", select=-gender) indep <- glm(Freq ~ right + left, data = women, family=poisson) mosaic(indep, main="Vision data: Independence (women)" ) quasi.indep <- glm(Freq ~ right + left + Diag(right, left), data = women, family = poisson) mosaic(quasi.indep, residuals_type="rstandard", gp=shading_Friendly, main="Quasi-Independence (women)" ) symmetry <- glm(Freq ~ Symm(right, left), data = women, family = poisson) mosaic(symmetry, residuals_type="rstandard", gp=shading_Friendly, main="Symmetry model (women)" ) quasi.symm <- glm(Freq ~ right + left + Symm(right, left), data = women, family = poisson) mosaic(quasi.symm, residuals_type="rstandard", gp=shading_Friendly, main="Quasi-Symmetry model (women)") 17

You can use anova() to compare a collection of nested models: # model comparisons: for *nested* models anova(indep, quasi.indep, quasi.symm, test="Chisq") anova(symmetry, quasi.symm, test="Chisq")

18

Logit models and logistic regression

Logit models are just a special case of logistic regression models in which all predictors are discrete. The only reasons for treating them specially are pedagogical: (a) every logit model for a binary response is equivalent to a loglinear model; (b) it is then a simple step to logistic regression models that include continuous predictors.

4.1

Logit models

Its time to have another look at the Berkeley data, from the perspective of logit models. In these models, with Admit as a response, we just specify the factors that Admit depends upon (Dept?, Gender?). 1. Fit a model in which Admit depends only on Dept. (The equivalent loglinear model is [AD][DG], which includes the association of Department and Gender.) 2. Fit the main eects model, Admit Dept + Gender. Plot the observed and tted logits under this model.

Tasks & questions

In SAS, logit models are most easily t with PROC CATMOD, and graphical displays of the SAS tted model most easily obtained with the CATPLOT macro. For this purpose, the option out=predict on the response statementroduces an output data set continaining tted values and standard errors. The statements below are contained in the le berkeley-logit.sas, which berkeleylogit.sas includes some additional plotting statements to make plots prettier. Eect of Dept only: %include catdata(berkeley); *-- logit (Admit) ~ Dept; proc catmod order=data data=berkeley; format dept dept. ; population dept gender; weight freq; response / out=predict; model admit = dept / ml noiter noprofile title="Model (AD, DG)"; run; Main eects of both Dept and Gender: proc catmod order=data data=berkeley; weight freq; response / out=predict; model admit = dept gender / ml noiter noprofile title="Model (AD, AG, DG)" ; run; Plotting the observed and tted values (on the logit scale) with the CATPLOT macro: %catplot(data=predict, class=gender, x=dept, type=FUNCTION, z=1.96); In R, the parallel analysis as a logit model is easiest if the data UCBAdmissions is rst be converted from a 2 2 6 table to a data.frame in frequency form. Models can then be R expressed for the variable Admit, using the Freq variable as a weight. The statements below are contained in the le berkeley-logit.R. berkeleylogit.R

Eect of Dept only: 19

library(car) # for Anova() data(UCBAdmissions) UCB.df <- as.data.frame(UCBAdmissions) berk.mod1 <- glm(Admit=="Admitted" ~ Dept, data=UCB.df, weights=UCB.df$Freq, family="binomial") Anova(berk.mod1, test="Wald") summary(berk.mod1) Adding a main eect of gender: berk.mod2 <- glm(Admit=="Admitted" ~ Dept+Gender, data=UCB.df, weights=UCB.df$Freq, family="binomial") Anova(berk.mod2, test="Wald") summary(berk.mod2) Plotting tted values. The effects package is explored in more detail later in this section. library(effects) ## load the effects package berk.eff2 <- allEffects(berk.mod2) plot(berk.eff2) plot(effect(Dept:Gender, berk.mod2), multiline=TRUE)

4.2
4.2.1

Logistic regression
Arthritis treatment data

In previous lectures, I used the Arthritis data, but always ignoring the covariate Age, so I could treat it as a contingency table, Sex Treatment Improve. Here, we will t logistic regression models, but only for a binary response, better = 0, improve = None 1, improve = Some,Marked
Tasks & questions

1. Fit a logistic model predicting Pr(better = 1) from Age alone. Make a graph to visualize how better changes with Age. 2. Add the eects of Sex and and Treatment to this model. Make sure you understand how these eects are parameterized in your model. Interpret the coecients for Age, Sex and Treatment on the probability of improvement. 3. Make a reasonable graph to visualize the eects of Age, Sex and Treatment on the probability of improvement. 4. Consider the possibilities there may be two-way interactions among Age, Sex and Treatment and the eect of Age may be non-linear (e.g., Age, Age2 ). Fit a model that tests for these additional terms over and above the main eects.

For SAS, the variable better has been dened in the data step that creates the arthrit data set. Use the option descending to model the probability better=1. In SAS 9.3, the SAS effectplot statementproduces reasonable plots either on the probability scale or logit scale. The statements below are contained in the le arthritis-logistic.sas. arthritislogistic.sas

Logistic regression on age alone:

20

%include catdata(arthrit); title Simple logistic regression on age; proc logistic data=arthrit descending; model better = age ; effectplot fit / obs(jitter(y=0.02)); effectplot fit / obs(jitter(y=0.02)) link; run;

* display on logit scale;

Fit the main eects model. Note the use of the ref= option to assign reference categories for Sex and Treat, and the output statemento obtain a data set for plotting with the meanplot macro. title main effects model; proc logistic data=arthrit descending; class sex (ref=last) treat (ref=first) / param=ref; model better = sex treat age ; output out=results p=prob l=lower u=upper xbeta=logit stdxbeta=selogit / alpha=.33; effectplot slicefit (sliceby=treat) / at(sex=all) clm alpha=0.33; effectplot interaction (x=treat sliceby=sex) / at(age=30 60) clm alpha=0.33 noobs; run; *-- or use meanplot macro; %meanplot(data=results, response=prob, class=age treat sex, pmean=no); To test potential high-order terms, you can use forward selection, starting with the main eects model. title screening for higher-order effects; proc logistic data=arthrit descending; class sex(ref=last) treat(ref=first) / param=ref; model better = age sex treat age | sex | treat @2 age*age / selection=forward slentry=0.1 /* be a bit liberal */ start=3 ; /* Start after all main effects */ effectplot slicefit (sliceby=treat) / at(sex=all)clm alpha=0.33; title2 Testing all interactions via forward selection; run; In R, use the Arthritis data in the vcd package. The variable Better must be calculated as shown below from Improved. Logistic regression models are t with glm(), and tted values R can be obtained with predict() for the glm object. The statements below are contained in the le arthritis-logistic.R. arthritislogistic.R

Logistic regression on age alone: library(vcd) library(car) data(Arthritis) # define Better Arthritis$Better <- Arthritis$Improved > None # simple linear regression arth.mod0 <- glm(Better ~ Age, data=Arthritis, family=binomial) 21

anova(arth.mod0) # plot, with +-1 SE plot(Better ~ Age, data=Arthritis, ylab="Prob (Better)") pred <- predict(arth.mod0, type="response", se.fit=TRUE) ord <-order(Arthritis$Age) lines(Arthritis$Age[ord], pred$fit[ord], lwd=3) upper <- pred$fit + pred$se.fit lower <- pred$fit - pred$se.fit lines(Arthritis$Age[ord], upper[ord], lty=2, col="blue") lines(Arthritis$Age[ord], lower[ord], lty=2, col="blue") Fitting the main eects model: # main effects model arth.mod1 <- glm(Better ~ Age + Sex + Treatment , data=Arthritis, family=binomial) Anova(arth.mod1) # same, using update() arth.mod1 <- update(arth.mod0, . ~ . + Sex + Treatment) Anova(arth.mod1) Forward selection: # forward selection from the main effects model step(arth.mod1, direction="forward", scope=.~ (Age+Sex+Treatment)^2 + Age^2) 4.2.2 Volunteering for psychology experiments

Cowles and Davis (1987) examined personality factors that might relate to whether an individual volunteered to participate in a psychology experiment. The predictors used here are scales of Neuroticism and Extraversion from the Eysenck personality inventory, plus Sex of the participant. There are 1421 observations. 1. Fit a main eects model predicting Pr (Volunteer = Yes) from Sex, Neuroticism and Extraversion. How well does this model t? 2. Fit a model that includes all two-way interactions among Sex, Neuroticism and Extraversion. How well does this model t? Which interaction terms appear not to contribute to prediction? 3. Fit a reduced model including only the important interaction(s) from the all two-way model. What do you conclude?

Tasks & questions

In SAS, the data are available in cowles.sas and some statements for this exercise in the le SAS cowles-logistic.sas. The statements below will get you started with the main eects model. cowlesNote the use of the descending option to model the event Volunteer=1 (yes). For logistic logistic.sas models, the lackfit option on the model statementgives the Hosmer-Lemeshow lack of t test %include catdata(cowles); *-- apply formats to Sex and Volunteer; data cowles; set cowles; 22

format Sex sex. Volunteer volun.; proc print data=cowles(obs=20); run; *-- main effects model; proc logistic data=cowles descending ; class Sex; model Volunteer = Sex Extraver Neurot / run;

lackfit ;

For other models, you can use | notation in the model statement here, @ forms all high-order terms up to a given degree. Thus, the all two-way model can be specied as *-- all two-way model; proc logistic data=cowles descending ; class Sex; model Volunteer = Sex | Extraver | Neurot @2 / run;

lackfit ;

For R, the data frame Cowles is contained in the car package, which also has an Anova() R function providing Type II and Type III tests, more useful than the sequential Type I tests provided by anova(). The R statements for this exercise are provided in the le cowleslogistic.R. For the main eects model, the commands are shown below. cowleslogistic.R

library(car) ## for Anova: type II tests data(Cowles) mod.cowles0 <- glm(volunteer ~ sex + neuroticism + extraversion, data=Cowles, family=binomial) summary(mod.cowles0) Anova(mod.cowles0)

4.2.3

Marijuana arrests data

In this exercise you will t and examine graphically a more realistic (and complex) model. The data concern police treatment of individuals arrested in Toronto for simple possession of small quantities of marijuana.4 Under these circumstances police have the option of releasing an arrestee with a summons to appear in court similar to a trac ticket; alternatively, the individual may be brought to the police station for questioning and possible indictment. The principal question of interest is whether and how the probability of release is inuenced by the subjects race, age, and other characteristics. Beyond this, there is also interest in whether race interacts with other variables. There are 5226 observations. The variables are: released Whether the arrestee was released with a summons: No or Yes colour The arrestees race: Black, White. year 1997 through 2002. Although numeric, we will treat year as a factor here. age in years sex Female, Male. employed No, Yes
This was part of a larger data set on racial proling, featured in a series of articles in the Toronto Star newspaper, December, 2002.
4

23

citizen No, Yes checks Number of police data bases (of previous arrests, previous convictions, parole status, etc. 6 in all) on which the arrestees name appeared 1. Fit a main eects logistic regression model predicting released from all the other variables. 2. Fit a model that includes all possible two-way interactions as well as all main eects. Which two-way interactions appear to be important? 3. Fit a reduced model that includes only the important two-way interactions as well as all main eects. Methods for model selection is a complex topic here, we will only use backward selection from the all two-way model to illustrate one simple approach.

Tasks & questions

For SAS, the data are read from the le arrests.sas, and some code for this exercise is in SAS the le Arrests-logistic.sas. Arrestslogistic.sas %include catdata(arrests); *-- main effects model; proc logistic data=arrests; class colour year sex citizen employed; model released = colour year age sex employed citizen checks; run; *-- all two-way model, with backward elimination; proc logistic data=arrests; class colour year sex citizen employed; run; model released = colour | year | age | sex | employed | citizen | checks@2 / selection=backward; run; For R, the Arrests data are contained in the effects package, and some statements for this R exercise are contained in the le Arrests-logistic.R. Note that, in model formulas, ~. stands Arrestsfor is modeled by all predictors; ~.^2 is a short-hand for all main eects plus two-way logistic.R terms. library(effects) # for Arrests data library(car) # for Anova() data(Arrests) Arrests$year <- as.factor(Arrests$year) # all main effects arrests.mod1 <- glm(released ~ ., family=binomial, data=Arrests) Anova(arrests.mod1) # all two-way effects arrests.mod2 <- glm(released ~ .^2, family=binomial, data=Arrests) Anova(arrests.mod2) To eliminate unnecessary two-way terms, try stepAIC(), that tries to minimize the AIC statistic. # backward selection, using AIC arrests.step <- stepAIC(arrests.mod2, direction="backward") anova(arrests.mod1, arrests.step, arrests.mod2, test="Chisq") 24

4.3

Eect plots

Eect plots are among the most useful graphical displays for understanding the pattern of eects in complex models. As explained in the lecture, the idea is relatively simple: to visualize a high-order term in a model, plot the predicted values (adjusted means) for that term, averaging over all other factors not included in that term. 4.3.1 Volunteering for psychology experiments
Tasks & questions

1. Construct an eects plot to visualize the Neuroticism * Extraversion interaction for the Cowles data. 2. Give a verbal description of how probability of volunteering changes with both neuroticism and extraversion.

In SAS 9.3, eect plots are easily provided by the effectplot statementThe program statements for this exercise are contained in the le cowles-eect.sas. cowles%include catdata(cowles); proc logistic data=cowles outest=parm descending ; class Sex; model Volunteer = Sex Extraver | Neurot / lackfit ; effectplot slicefit(x=Extraver sliceby=Neurot) / at(sex=1.5) noobs; effectplot slicefit(x=Neurot sliceby=Extraver) / at(sex=1.5) noobs; effectplot contour(x=Neurot y=Extraver) / at(sex=1.5) noobs; run;
eect.sas

In R, eect plots for a wide class of linear models are provided in the effects package. The function allEffects() creates the predicted eect values for all high-order terms in the model. These can then be plotted with the plot() method. The statements below are R contained in the le cowles-eect.R. They show two dierent forms of eect plots for the cowleseect.R neuroticism:extraversion eect. library(effects) ## load the effects package data(Cowles) mod.cowles <- glm(volunteer ~ sex + neuroticism*extraversion, data=Cowles, family=binomial) summary(mod.cowles) eff.cowles <- allEffects(mod.cowles, xlevels=list(neuroticism=seq(0, 24, 6), extraversion=seq(0, 24, 8))) #-- separate panels plot(eff.cowles, neuroticism:extraversion, ylab="Prob(Volunteer)", ticks=list(at=c(.1,.25,.5,.75,.9)), layout=c(4,1)) #-- multiline plot plot(eff.cowles, neuroticism:extraversion, multiline=TRUE, ylab="Prob(Volunteer)", key.args=list(x = .8, y = .9)) 4.3.2 Marijuana arrests data

Here we examine eect plots for high-order terms in one model for released in the Arrests data. This model allows for interactions of color with both year and age. As mentioned above, these plots are easy to do with the effects package in R, but more dicult in SAS. 25

Tasks & questions

1. Fit the model that includes terms for colour*year and colour*age, as well as other variables as main eects. 2. Obtain an eect plot for the 3-way term colour:year:age (not included in the model). Interpret this plot in relation to how the probability of release varies with both Year and Age. 3. Obtain separate eect plots for the 2-way terms colour:age and colour:year In SAS 9.3, we can again use the effectplot statement Note how the various options control the details. The statements below are contained in the le Arrests-eect.sas.

SAS
Arrests-

proc logistic data=arrests descending; eect.sas class colour year sex citizen employed; model released = colour|year colour|age sex employed citizen checks / clodds=wald; effectplot interaction (x=year sliceby=colour) / clm alpha=0.33 noobs yrange=clip; effectplot slicefit (x=age sliceby=colour) / clm alpha=0.33 obs(fringe jitter) yrange=(.7, 1); run; The statements below are contained in the le Arrests-eects.R. Fit the model: library(effects) data(Arrests) Arrests$year <- as.factor(Arrests$year) arrests.mod <- glm(released ~ employed + citizen + checks + colour*year + colour*age, family=binomial, data=Arrests) Three-way eect plot: plot(effect("colour:year:age", arrests.mod, xlevels=list(age=15:45)), multiline=TRUE, ylab="Probability(released)", rug=FALSE) Two-way eect plots. Note the dierence between the style of plot with multiline=TRUE and multiline=FALSE. plot(effect("colour:age", arrests.mod), multiline=TRUE, ylab="Probability(released)") plot(effect("colour:age", arrests.mod), multiline=FALSE, ylab="Probability(released)") For complex models, a convenient feature of the effects package is the ability to obtain the necessary statistics for all eects simultaneously, and plot various terms by selection from a menu. arrests.effects <- allEffects(arrests.mod, xlevels=list(age=seq(15,45,5))) plot(arrests.effects, ylab="Probability(released)") R
Arrestseects.R

26

4.4

Diagnostic plots

In this exercise, the goal is to explore diagnostic plots for generalized linear models, which are simple extensions of similar plots for ordinary linear models (regression and ANOVA). We return to the Berkeley data one more time. 1. Fit the model [AD][GD] of conditional independence, as before. 2. Construct plots to show the distribution of residuals from this model and to identify inuential cells in terms of residual and leverage. 3. Interpret what you see in these plots in relation to other plots of these data in previous exercises.

Tasks & questions

In SAS, you can use the inflglim and halfnorm macros, as well as the ODS Graphics facility SAS in SAS 9.2+ to obtain a wide variety of diagnostic plots. See berkeley-diag.sas. berkeleydiag.sas

Using the inflglim and halfnorm macros. For these plots, begin by constructing a character variable that joins the factor levels into a single cell variable for labels in the plots. %include catdata(berkeley); *-- make a cell ID variable, joining factors; data berkeley; set berkeley; cell = trim(put(dept,dept.)) || gender || trim(put(admit,yn.)); run; *-- Conditional independence; proc genmod data=berkeley; class dept gender admit; model freq = dept|gender dept|admit / dist=poisson obstats residuals; ods output obstats=obstats; Use the inflglim macro to produce inuence plots for this model. %inflglim(data=berkeley, class=dept gender admit, resp=freq, model=dept|gender dept|admit, dist=poisson, id=cell, gx=hat, gy=stresdev); %inflglim(data=berkeley, class=dept gender admit, resp=freq, model=dept|gender dept|admit, dist=poisson, id=cell, gx=hat, gy=difdev); Use the halfnorm macro to produce a half-normal residual plot for this model, with a simulated 95% envelope for the mean response in the model. %halfnorm(data=berkeley, class=dept gender admit, resp=freq, model=dept|gender dept|admit, dist=poisson, id=cell); Using ODS graphics with plot=all. This only works in SAS 9.2+, and the plots are all index plots variable plotted against observation (cell) number. proc genmod data=berkeley plots=all; class dept gender admit; model freq = dept|gender dept|admit / dist=poisson; run; 27

In R, some of these plots can be obtained from the plot() method for glm objects. An inuence plot is provided by influencePlot() in car package. See berkeley-diag.R. berkeleydiag.R

library(car) berkeley <- as.data.frame(UCBAdmissions) cellID <- paste(berkeley$Dept, substr(berkeley$Gender,1,1), -, substr(berkeley$Admit,1,3), sep="") rownames(berkeley) <- cellID berk.mod <- glm(Freq ~ Dept * (Gender+Admit), data=berkeley, family="poisson") influencePlot(berk.mod, labels=cellID, id.n=3) plot(berk.mod)

28

5
5.1

Polytomous response models

Proportional odds model
Arthritis data

5.1.1

In this exercise you will try to t a proportional odds model to the Arthritis data, modeling the 3-category response Improve (None, Some, or Marked) and using Sex, Treatment and Age as predictors. 1. Fit a main eects model predicting improve from Sex, Treatment and Age. Note that both Sex and Treatment are binary factors (CLASS variables in SAS), so the interpretation of the coecients depends on how they are coded. 2. How well does the proportional odds model t for these data?5 3. Interpret the coecients for Sex, Treatment and Age in this model. For example, how would you interpret the coecient for Sex in relation to the odds of a better outcome? 4. Obtain and plot predicted (tted) values on either the probability scale or the logit scale. Interpret this plot in relation to the numerical results (coecients, odds ratios, etc.)

Tasks & questions

For SAS, the steps are illustrated in the lecture notes, Proportional odds model: Fitting and plotting. Below, I show just the steps to t the model, and obtain an output data set SAS with the tted probabilities and logits. Note that Sex is coded so that Male is the reference category, and Treat so that Placebo is the reference category. The additional steps to produce nicely formatted plots are contained in the le arthritis-propodds.sas. arthritispropodds.sas

title Logistic Regression: Proportional Odds Model; %include catdata(arthrit); proc logistic data=arthrit descending; class sex (ref=last) treat (ref=first) / param=ref; model improve = sex treat age ; output out=results p=prob l=lower u=upper xbeta=logit stdxbeta=selogit / alpha=.33; proc print data=results(obs=6); id id treat sex; var improve _level_ prob lower upper logit; format prob lower upper logit selogit 6.3; run; In SAS 9.2+ it is easier to plot results using ODS graphics and the effectplot statement The code below is contained in arthritis-propodds-ods.sas arthritispropodds-

proc logistic data=arthrit descending ; ods.sas class sex (ref=last) treat (ref=first) / param=ref; model improve = sex treat age / clodds=wald expb; effectplot slicefit(sliceby=improve plotby=Treat) / at(sex=all) clm alpha=0.33 ; effectplot interaction(x=Treat sliceby=improve) / at(sex=all) noobs clm alpha=0.33; run; In R, we can t the proportional odds model with polr() in the MASS package. There is
Caution: The score test for the proportional odds model is known to be anti-conservative, i.e., it yields p-values that are too small in many cases.
5

29

no simple equivalent for the score test of the proportional odds assumption. The statements below are contained in the le arthritis-propodds.R. arthritispropodds.R

library(vcd) library(car)

# for Anova()

arth.polr <- polr(Improved ~ Sex + Treatment + Age, data=Arthritis) summary(arth.polr) Anova(arth.polr) # Type II tests In R it is easy to obtain eect plots showing the tted probabilities in relation to any of the terms in the model (or even terms not in the model). library(effects) arth.effects <- allEffects(arth.polr, xlevels=list(age=seq(15,45,5)) ) plot(arth.effects) # visualize all main effects # plot terms not in model plot(effect("Sex:Age", arth.polr)) plot(effect("Treatment:Age", arth.polr))

5.2

Nested dichotomies

This exercise concerns data on womens labour-force participation collected at York University in a 1977 survey of the Canadian population. The three-level response has categories Not working outside the home (n = 155), working Part-time (n = 42) and working Full-time (n = 66). The predictors are Husbands income ($1000s), Presence of children in the household, and Region of Canada. Region turns out not to be an important predictor, so we will ignore it here. 1. The rst question is whether the proportional odds model provides a reasonable account of these data if so, it would give a simple description. Fit the proportional odds model, and test whether the proportional odds assumption is tenable (only in SAS). 2. Fit separate logistic models for the two nested dichotomies: Working (full- or part-time) vs. Not Working; and Working Full-time vs. Part-time. 3. Interpret the coecients in the two models. What do they say about the eect of Husbands income and Children on the probabilities for the two dichotomies? 4. Plot the tted probabilities for the 3 categories of the response in relation to Husbands income and Children.

Tasks & questions

For SAS, the data are contained in the data set wlfpart. The 3-level response here is labor, SAS with values 1 (Full-time), 2 (Part-time), 3 (Not working). When the data are read in, two dichotomous responses are created: working = labor < 3, and fulltime = labor=1 (but just for working women. The statements below (and more) are contained in the le wlfnested.sas. wlf-nested.sas Fitting the proportional odds model:

30

%include catdata(wlfpart); proc logistic data=wlfpart nosimple descending; model labor = husinc children ; title2 Proportional Odds Model for Fulltime/Parttime/NotWorking; run; Fitting and plotting the models for the nested dichotomies. Note how an ODS OUTPUT statementan be used to capture the global test statistics. See the wlf-nested.sas le for how to wlf-nested.sas combine these to give overall tests for the 3-level nested dichotomy model. proc logistic data=wlfpart nosimple descending; model working = husinc children ; output out=resultw p=predict; ods output GlobalTests=gtests1; title Nested Dichotomies; proc plot; plot predict * husinc = children; proc logistic data=wlfpart nosimple descending; model fulltime = husinc children ; output out=resultf p=predict; ods output GlobalTests=gtests2; proc plot; plot predict * husinc = children; run; quit; In R, the data are in the data frame Womenlf in the car package. The 3-level response is called R partic here, an unordered factor with values fulltime, not.work and parttime. One easy way to create the two dichotomous responses is with the recode() function in car package. The le wlf-nested.R contains the statements below, and also shows how to produce a plot of wlf-nested.R the tted values. library(car) data(Womenlf) attach(Womenlf) working <- recode(partic, " not.work = no; else = yes ") fulltime <- recode (partic, " fulltime = yes; parttime = no; not.work = NA") Fitting nested dichotomies models: We use treatment contrasts for children, so the coecient represents the eect for having children. contrasts(children)<- contr.treatment mod.working <- glm(working ~ hincome + children, family=binomial) summary(mod.working) mod.fulltime <- glm(fulltime ~ hincome + children, family=binomial) summary(mod.fulltime) Anova(mod.working) Anova(mod.fulltime) Obtaining predicted values for plotting: The general idea is to create a data frame with the combinations of the predictor values you want to see, and use the predict() method to get 31

the tted values. For binomial glm() models, type=response gives tted values on the probability scale.6 # get fitted values for both sub-models predictors <- expand.grid(hincome=1:45, children=c(absent, present)) p.work <- predict(mod.working, predictors, type=response) p.fulltime <- predict(mod.fulltime, predictors, type=response) One slight complication: Because the model for the fulltime dichotomy is conditional on working, you need to do some calculations to obtain the unconditional probabilities for each of the three response categories. # calculate unconditional probs for the three response categories p.full <- p.work * p.fulltime p.part <- p.work * (1 - p.fulltime) p.not <- 1 - p.work

5.3

Generalized logits

The generalized (or multinomial) logit model is an alternative to using nested dichotomies for a polytomous reponse. One advantage is that you can t this as a single model, without resorting to the special tricks used for nested dichotomies. In using the generalized logit model, you should take care to specify the baseline response category, so you understand how to interpret the tted coecients. As always, plots of the tted model (probabilities or logits) facilitate interpretation. In this exercise, we continue with the data on womens labor participation from Section 5.2, but now t a generalized logit model. 1. Fit the generalized logit model, using Not working as the baseline response category. 2. Examine the coecients for Husbands income and Children in the comparisons of Parttime and Full-time vs. Not working. On which comparison to the predictors show a greater impact? 3. Plot the tted probabilities of all three response categories over the range of Husbabds income for both levels of Children. Compare this plot to the analogous plot for the nested dichotomies model. In SAS, you can t the generalized logit model for the polytomous response labor simply by specifying the option link=glogit on the model statement Use the output statemento obtain tted values, both on the probability scale p=predict and on the logit scale xbeta=logit. *-- Fit generalized logit model; proc logistic data=wlfpart; model labor = husinc children / link=glogit; output out=results p=predict xbeta=logit;

Tasks & questions

SAS

In the output data set results, the three response categories are identied by a variable _level_. The simplest plot of the tted probabilities is produced with PROC PLOT as shown below. See the le wlf-glogit.sas for details on plotting the tted results with PROC GPLOT and wlf-glogit.sas better labels for the curves.
6

The default, type=link, gives tted values on the logit scale.

32

proc sort data=results; by children husinc _level_; *-- simple plot; proc plot data=results; plot predict * husinc = _level_ ; by children; For R, it is helpful to rst turn partic into an ordered factor, so that Not working is the baseline category. The multinomial logit model cannot be t directly with glm(); it can be R t using the multinom() function in the nnet package, which also requires the MASS library. See the le wlf-glogit.R for more details on plotting the tted results. wlf-glogit.R Fitting the multinomial logit model: library(car) data(Womenlf) attach(Womenlf) participation <- ordered(partic, levels=c(not.work, parttime, fulltime)) library(nnet) library(MASS) mod.multinom <- multinom(participation ~ hincome + children) summary(mod.multinom, cor=F, Wald=T) Anova(mod.multinom) Obtaining predicted values for plotting: Here, we get tted values on the probability scale with type=probs. predictors <- expand.grid(hincome=1:45, children=c(absent, present)) p.fit <- predict(mod.multinom, predictors, type=probs)

33